Imagine an AI that could coach you, offer constructive criticism, and help you improve your work. That's the promise of AI feedback systems. But can AI truly understand nuanced concepts like empathy and effective communication to provide helpful feedback? New research explores this question by examining how different guidelines, or "constitutions," influence an AI's ability to critique and improve medical conversations. Researchers tested four different sets of guidelines, ranging from detailed best practices for patient-centered communication to a simple instruction to "act like a good doctor." They then used these guidelines to train an AI to provide feedback on simulated doctor-patient dialogues. The results? While detailed guidelines improved the AI's feedback on emotionally sensitive aspects of the conversation, like showing empathy and shared decision-making, they didn't improve the AI's ability to coach on more practical aspects, such as information gathering. This suggests that while AI can learn to recognize and reward certain behaviors, it may struggle with more complex reasoning tasks. The study also highlighted the challenge of aligning AI feedback with human preferences. Even when the AI successfully followed its guidelines, human evaluators sometimes disliked the results, finding the AI's responses verbose or unnatural. This points to the ongoing need to refine how we train and evaluate AI feedback systems, ensuring they not only follow rules but also genuinely understand and respond to human needs. This research is a critical step towards developing truly helpful AI assistants, but it also reveals the complexities of teaching machines to understand and respond to the nuances of human communication. The future of AI feedback hinges on bridging this gap, allowing AI to provide not just rule-based critiques but genuinely insightful and helpful guidance.